Research And Implement Of Instance Segmentation Of Dairy Goat Image Based On Transformer

The quality of image segmentation plays a decisive role in the analysis and understanding of images of livestock.By high-quality instance segmentation of dairy goat images,accurate monitoring of dairy goat image information can be achieved,which facilitates the precision and intelligence development of dairy goat farming industry.This thesis uses the dairy goat images acquired from the animal husbandry teaching and experimental base of Northwest A&F University as the research object,and implements the instance segmentation of dairy goat images based on Transformer.The main research contents and conclusions of this thesis are as follows:(1)Construction of dairy goat image instance segmentation dataset.To address the problem of lacking public datasets of dairy goat images that are easy to obtain on the Internet,monitoring videos of dairy goat activities were collected at different time periods in the dairy goat farm.After collection,FFmpeg software was used to extract key frames from the videos and preprocess them to obtain high-quality and uniformly sized dairy goat images.Then,Label Me software was used to annotate the images,and two types of augmentation methods,geometric and color transformation,were used to expand the dataset.Finally,a high-quality dairy goat image instance segmentation dataset was obtained.(2)Instance segmentation of dairy goat image based on Transformer.Based on DETR,a query-based instance segmentation method using Transformer is proposed,which unifies the category,position,and mask of the target object into a query embedding,enabling simultaneous classification,bounding box regression,and mask prediction tasks.A unified query representation module is constructed to replace the prediction head of the DETR network structure,achieving end-to-end dairy goat image instance segmentation.(3)Improvement of Transformer module based on deformable attention.To address the problems of high computational complexity,training time,and inefficient use of feature information when processing image data with Transformer,a deformable attention mechanism is proposed to improve the computational efficiency of Transformer.In addition,a multi-scale deformable attention module is introduced to construct multi-scale features,and the backbone network is replaced with Res2 Net,both of which improve the multi-scale feature extraction ability of the network,thereby improving the accuracy of dairy goat image instance segmentation.Experimental results show that the average segmentation accuracy of this method reaches 69.20%,with good segmentation performance.(4)Design and implementation of dairy goat image instance segmentation system.In order to facilitate the improvement research of dairy goat image monitoring technology and other downstream tasks and the practice of precision breeding of dairy goats,a dairy goat image instance segmentation system based on B/S architecture is designed and implemented based on previous research results.Through testing,the system satisfies the demand analysis requirements,operates stably,and has good segmentation performance,being of certain practical value.